Asset Mgmt Control Plane Model Calls | Moring
Model calls are the requests agents send to language models, and governing them at the control plane means every call is authenticated under the agent's identity, evaluated against policy, routed to approved models, budgeted, and logged on one path. Architecture research states the rule directly: no consumer calls a model directly. For asset management firms, this is where data handling, model choice, and spend become enforceable rather than promised.
Short answer
Model calls are the requests agents send to language models, and governing them at the control plane means every call is authenticated under the agent's identity, evaluated against policy, routed to approved models, budgeted, and logged on one path. Architecture research states the rule directly: no consumer calls a model directly. For asset management firms, this is where data handling, model choice, and spend become enforceable rather than promised.
How it works
Model calls are the most frequent governed event in an agentic estate, which makes them the natural place to test whether governance is real. Architecture research on the control plane states the founding rule plainly: no consumer calls a model or an enterprise capability directly; every agent-to-model call transits the governing layer. Commentary on the same architecture adds the operational questions this answers that runtimes were never designed to ask: which agent called which model, under whose authorization, with what data, at what cost, and whether the endpoint was cleared to receive it.
The control plane treats a model call as a governed transaction with a subject and an object. The subject is the agent's scoped identity, verified on each request; the object is the model endpoint, selected from the firm's approved set. Definition research on the category describes the layer as unifying connection, identity, policy enforcement, and observability so every prompt, response, and tool call flows through a single controlled path. For an asset manager, this means a research workflow's calls to external models are checked for data classification before transit, and a reporting workflow's calls are routed to the model tier policy assigns for that sensitivity.
Policy on model calls is more than access control. Guidance on model governance controls describes the inventory discipline: a continuously refreshed account of which models operate where, rather than a spreadsheet updated quarterly, because policy can only bind models the firm knows it is using. The control plane makes the inventory a byproduct of routing, since every call reveals the model in use. Buyer research for financial services adds the spend dimension, noting teams connect agents to enterprise systems and consume model capacity with or without governance; per-call attribution at the plane turns that consumption into an accountable budget line.
Runtime evaluation matters because agent behavior is not fixed at deployment. Agents plan multi-step workflows, and the sensitivity of successive model calls varies within a single run: early retrieval steps may handle public research while later drafting steps process client-specific information. Enterprise policy architecture research describes decisions returned per request, allowing, denying, or editing the call, so a prompt carrying restricted data can be transformed or blocked while the same workflow's other calls proceed. Static per-agent rules cannot make that distinction; the enforcement point can.
Ownership of the enforcement point is the strategic question. Research on control plane ownership argues the enterprise must hold it, because delegating enforcement to an external provider or platform hands the decisive control to the ecosystem being governed. For asset management firms bound by supervisory expectations for model risk management that now translate to the agent layer, and subject to EU AI Act traceability obligations on autonomous systems, holding the model-call enforcement point is what keeps the compliance record internal and complete.
The evidence trail is the final product. Every model call recorded with identity, policy decision, model, tokens, and cost produces the audit granularity that examinations, incident investigations, and cost attribution all draw from. Governance research emphasizes that controls reduce risk by observing, enforcing, and leaving evidence; the model-call path is where all three happen at the highest frequency in the estate.
Frequently asked questions
What does governing model calls at the control plane involve?
Authenticating each call under the agent's scoped identity, evaluating it against policy at runtime, routing it to approved model endpoints, applying budgets, and logging identity, decision, model, tokens, and cost on one path. Architecture research states the rule that no consumer calls a model directly, so every agent-to-model call transits the governing layer. For asset management, this makes data classification, model choice, and spend enforceable per call rather than promised per deployment.
Why is runtime evaluation of model calls necessary for asset managers?
Because agent behavior varies within a single workflow: early retrieval steps may handle public research while later drafting steps process client-specific information. Enterprise policy architecture research describes per-request decisions that allow, deny, or edit the call, so a prompt carrying restricted data can be transformed or blocked while other calls proceed. Static per-agent rules cannot make that distinction. Supervisory expectations for model risk management extend to the agent layer, and EU AI Act traceability obligations cover autonomous systems, so per-call records at the control plane matter.
How do model calls relate to the approved model inventory?
Every call transiting the control plane reveals which model is in use, making the inventory continuously refreshed rather than a quarterly spreadsheet, which governance research identifies as the foundational control. Routing policy then restricts calls to the approved set, so an unapproved model cannot enter production through a developer's convenience. Asset management firms use this to keep model risk documentation aligned with the models actually operating, closing the gap between what is documented and what agents call.
Why should the firm own the model-call enforcement point?
Because research on control plane ownership concludes that enforcement inside an external provider or platform delegates the decisive control to the same external ecosystem the firm is trying to govern. Owning the point keeps policy, identity, and evidence internal and portable across model vendors. Buyer research for financial services notes teams consume model capacity with or without governance, so attribution and spend control also depend on the firm-held path. Asset management supervision and EU AI Act traceability both presume records the firm controls.